Greetings,

The open source LLM scene is on fire. Just this week, one model dethroned Llama 2, another achieved half of GPT-4s coding performance while being a fraction of the size. And that’s just a small selection of the mountains of new projects that are released every day.

• Autonomous AI agents collaborate, design, code, test, and document in this virtual software company – ChatDev completes the whole software development process in under seven minutes for less than $1. - GitHub
  
  
• Protect your LLM-powered apps from prompt injection – PIPE is a detailed guide on what prompt injection is, why it’s an important problem, and how to mitigate it. - GitHub
  
  
• Falcon 180B is the new biggest open-source LLM, with more than double the amount of parameters of Llama 2 – it’s now also the highest scoring open model, beating Llama 2 by 1.39 points. - Hugging Face
  
  
• Tiny models are on the rise – Refact LLM is a 1.6 billion parameter code model achieving state-of-the-art performance and 32% on the HumanEval programming benchmark (for reference, base GPT-4 achieved 67%). - Refact
  
  
• Weird behavior during model training leads to the realization that LLMs can rapidly memorize data after just one “look” at it, prompting the hypothesis that maybe the standards for training and using LLMs need some re-thinking. - Fast.ai
  
  
• Google researchers strapped an LLM to OSS-Fuzz, the #1 suite for automated vulnerability discovery for open source projects – big step towards “a future of personalized vulnerability detection with little manual effort from developers.” - Google
  
  
• Manned fighter jets cost up to $143 million, AI-controlled XQ-58A Valkyrie drones only $3 million – the US Army wants to spend $6 billion to build a whole fleet of them. - Engadget

----

Found this helpful? Forward this email to a colleague.

Was this email forwarded to you? Sign up here to stay on top of AI news.